表面增強拉曼散射(SERS)技術憑藉其指紋光譜特性與極高的靈敏度，已成為精準醫療與多目標生物感測的重要工具。然而，現有的 SERS 檢測常受限於大型儀器的成本與體積，且在複雜混合樣本的定量分析上，傳統方法往往難以克服光譜重疊與基線漂移的干擾。本研究針對上述限制，建立可於邊緣裝置部署之定量分析模型與其驗證流程。在演算法層面，本研究提出寬主幹卷積神經網路架構(WideTrunkCNN)，針對多種 SERS 訊號分子之合成混合光譜進行訊號分子辨識與相對組成比例估計，並嵌入一維卷積區塊注意力模組(CBAM)以提供可映射回波數座標之注意力權重，供模型判斷依據之初步檢視。經架構診斷與改良確立網路架構後，再依真實配方特性調整訓練資料之合成分布，最終模型於獨立合成測試集(n = 1，000)上達到決定係數 R² = 0.9347、RMSE = 0.0392。純成分光譜經 airPLS 基線校正以抑制螢光背景後，作為合成混合光譜之基底。模型則以 CNN 進行自動化特徵萃取，並以主成分分析檢視多種訊號分子純光譜之可區別程度。為初步檢驗模型於真實混合物量測情境之適用性，本研究另以多筆獨立配製之真實混合物樣本進行驗證，模型於未經事後校正之情況下達 R²=0.64。於邊緣部署方面，本研究將模型匯出為 ONNX 格式並於微型電腦( Raspberry Pi )完成實機推論驗證，採用ONNX Runtime 可較 PyTorch 加速 8至13 倍。本研究涵蓋多重 SERS 定量之演算法開發、真實樣本驗證與邊緣推論可行性評估，在未來的研究上將進行真實樣本規模擴增與跨批次驗證等項目，進一步最佳化使本研究能推展至更多元之檢測與實際臨床應用。;Surface-enhanced Raman scattering (SERS) offers high sensitivity and molecular fingerprinting capability for precision medicine and multiplexed biosensing. However， quantitative analysis of complex SERS mixtures remains challenging because of spectral overlap and baseline drift， while conventional Raman systems are often costly and bulky. This study developed an edge-deployable quantitative analysis model and validation framework for multiplex SERS spectra. A wide-trunk convolutional neural network (WideTrunkCNN) was developed to identify multiple SERS reporter molecules and estimate their relative compositional proportions from synthetic mixture spectra. A one-dimensional Convolutional Block Attention Module (1D-CBAM) was incorporated to provide wavenumber-resolved attention weights for preliminary interpretation of model predictions. Pure-component spectra were processed using airPLS baseline correction and used to generate synthetic mixture spectra. Principal component analysis was also performed to examine the separability of the reporter spectrum. After refining the architecture and adjusting the synthetic training distribution to match experimentally prepared mixture characteristics， the final model achieved R² = 0.9347 and RMSE = 0.0392 on an independent synthetic test set (n = 1，000). Several independently prepared real-mixture samples were further evaluated， yielding R² = 0.64 without post-hoc calibration. For edge deployment， the model was exported to ONNX and tested on a Raspberry Pi， where ONNX Runtime achieved an approximately 8–13-fold inference speedup compared with PyTorch. Overall， this study demonstrates the feasibility of combining deep learning-based multiplex SERS quantification with edge inference. Future work will focus on larger-scale real-sample validation， cross-batch evaluation， and further optimization for broader detection and potential clinical applications.
